An unsupervised learning method for intelligent health monitoring of photovoltaic systems

Abstract

The present invention discloses an unsupervised learning method for intelligent health monitoring of photovoltaic systems, which belongs to abnormality detection of photovoltaic system panels. By collecting electroluminescent images of photovoltaic panels, that is, EL images of photovoltaic panels, an abnormality detection training platform is established, and the EL images of photovoltaic panels are used as input data. With the help of the EL images of photovoltaic panels, an abnormality detection network of the abnormality training platform is enabled. The trained abnormality detection network is used for image detection of photovoltaic panel arrays. The detection result is an abnormality score. The threshold of the abnormality score can be set by itself. The EL image of the photovoltaic panel is identified by the threshold to determine whether the current EL image is faulty. The present invention adopts the above method, and through a feature extraction network based on multi-head linear attention and an unsupervised abnormality detection framework based on scale learning, it can make up for the defect of insufficient deep information extraction ability of the current recognition method, thereby realizing abnormality detection of the EL image of the photovoltaic panel.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic system panel anomaly detection, and in particular to an unsupervised learning method for intelligent health monitoring of photovoltaic systems. Background Art

[0002] With the transformation of the global energy structure and the urgent need for renewable energy, photovoltaic technology has attracted much attention due to its clean and efficient characteristics. In the photovoltaic power generation system, as a key component, the performance of photovoltaic panels directly affects the overall power generation efficiency and operational stability of the system. However, in the actual use environment, photovoltaic panels are often affected by various complex factors, including local shading, material aging and other challenges. These factors may cause output power fluctuations, bringing potential operational risks to various equipment and systems that rely on photovoltaic power generation. Therefore, it is particularly important to detect anomalies in photovoltaic panels. Traditional deep learning models rely on a large number of labeled samples for training. However, in the actual application of the photovoltaic industry, the acquisition of a large number of labeled panel images requires professional equipment and expert knowledge, which will undoubtedly consume a lot of manpower and material resources. Therefore, some deep models based on data reconstruction, such as autoencoders and adversarial generative networks, are used for unsupervised anomaly detection of photovoltaic panels and have become the mainstream in the field of unsupervised anomaly detection. Although these methods are simple in principle, they use the deviation of the distribution of test samples from normal samples for anomaly detection, and to a certain extent get rid of the need for a large number of labeled samples. However, these anomaly detection methods have some defects in practical applications: First, when the data dimension is high, this reconstruction process consumes a lot of computing resources. Second, in the training process of the reconstruction-based anomaly detection model, the goal is to minimize the reconstruction error, which forces the model to focus on fine-grained information, resulting in noise being misjudged as useful information and thus learned by the model, making the model less resistant to noise; third, the focus on fine-grained information will make the model's ability to extract global information poor and unable to pay attention to the deep features of the data. Summary of the invention

[0003] The purpose of the present invention is to provide an unsupervised learning method for intelligent health monitoring of photovoltaic systems. In order to solve the problems in the background technology, for the first problem, a linear attention mechanism is used in the feature extraction process, which greatly reduces the computational requirements of this stage, and the core anomaly detection mechanism does not need to reconstruct the input data, but strategically generates supervision signals, thereby further reducing the computational burden. For the second problem, the scale learning mechanism gradually dilutes the influence of noise during the random sampling of input features, and the scale distribution alignment learning mechanism focuses on the relative ordering between feature subvectors, weakens the focus on the absolute value of the feature, and thus reduces the influence of noise on the model. For the third problem, a multi-head attention mechanism is used to fully integrate the global information from the photovoltaic EL image. In the scale distribution alignment learning mechanism, the arrangement information distribution of different subspace features in normal samples is learned, and these information encapsulates rich advanced features, thereby making up for the defect of insufficient deep information extraction ability, thereby realizing the effective extraction of global information and deep features in the photovoltaic panel EL image.

[0004] To achieve the above objectives, an unsupervised learning method for intelligent health monitoring of photovoltaic systems is proposed, which includes the following steps:

[0005] S1: Collect image data, establish a training platform for anomaly detection network, collect EL images of monocrystalline and polycrystalline photovoltaic panels under constant lighting conditions and constant load, pre-process the collected EL images, and divide the image data set into training set and test set x i represents an EL image sample, y i is a label, which indicates a faulty sample or a healthy sample. In an unsupervised scenario, the training set consists of healthy samples, and the test set includes both faulty samples and healthy samples.

[0006] The training platform of the anomaly detection network includes a multi-head linear attention-based feature extraction network and a scale-learning-based unsupervised anomaly detection framework;

[0007] S2: Input the training set into the feature extraction network based on multi-head linear attention to model the deep features of photovoltaic EL images and extract image features;

[0008] S3: The image features extracted in step S2 are converted into one-dimensional vectors and then input into the unsupervised anomaly detection framework based on scale learning. Through random sampling, feature conversion, and scale calculation methods, effective supervision signals are generated for distribution alignment learning to obtain a trained anomaly detection network.

[0009] S4, input the test set into the trained anomaly detection network, calculate the anomaly score, and perform unsupervised anomaly detection on the test set images by setting the anomaly judgment threshold;

[0010] S5. Deploy the trained anomaly detection network in a real working environment, and use the image data generated by actual working conditions for transfer learning fine-tuning. In engineering practice, verify the fault recognition effect and robustness of the trained model for photovoltaic images under different states.

[0011] Preferably, in step S1, data enhancement technology is used to construct a training set and a test set.

[0012] Preferably, in step S2, the feature extraction network based on multi-head linear attention includes five stages, namely, an input stage and four subsequent feature extraction stages, the input stage is a convolutional layer module and a depthwise separable convolutional module, in the first and second feature extraction stages, an inverted linear bottleneck layer module is provided, the step size of the inverted linear bottleneck layer module in each stage is set to 2, and in the third and fourth feature extraction stages, a depthwise separable convolutional module and an EfficientViT module are provided;

[0013] In the EfficientViT module, there is a multi-scale linear attention module for global information extraction, as well as a feedforward layer and a deep convolution layer for local information extraction. The input is linearly projected to obtain the Q / K / V matrix, and the Q / K / V matrix is ​​aggregated with a lightweight small-kernel convolution to obtain a multi-scale matrix. The multi-scale matrix is ​​linearly focused on by ReLU, and the output is connected to the final linear projection layer for feature fusion. The linear attention calculation formula is as follows:

[0014]

[0015] In the result of the above formula, when calculating linear attention, the calculation and

[0016] The results of the first, third and fourth feature extraction stages are upsampled, the number of channels and spatial dimensions are adjusted to remain consistent, and feature fusion is performed. The feature fusion results are output through an inverted linear bottleneck layer, and the output result is the image feature.

[0017] Preferably, in step S3, the process of performing distribution alignment learning is as follows;

[0018] Create a scale-based supervision signal, perform random sampling on the input one-dimensional vector several times, and obtain the randomly sampled features for a specific feature subspace S i , the features of which Input them into the feature conversion function T and the scale label calculation function G respectively, and obtain the data sample O output by the feature conversion function T and the supervision label γ output by the scale label calculation function G;

[0019] The feature subspace S i Features in The input is converted into n-dimensional data by the feature conversion function T. The feature conversion function T is defined as a simple feedforward layer with random initialization. Each feature subspace corresponds to a transformation layer, and the formula is as follows:

[0020]

[0021] In the above formula, represents the v-dimensional subvector, X v ∈R n×v Represents the weight matrix, b∈R n represents the bias term;

[0022] For m randomly sampled subvectors of a data instance d, their transformation is represented by the matrix W∈R m×n Indicates that W is regarded as a single data sample for scale learning, and the formula for data sample O is as follows:

[0023]

[0024] In the above formula, r is the total number of columns of data sample O;

[0025] The characteristics Input into the scale label calculation function G, the formula is as follows:

[0026]

[0027] In the above formula, n is the representation dimension of the sub-vector, α k is the weight of the kth feature, γ is the amplification factor, α k The formula is as follows:

[0028]

[0029] In the above formula, the weight calculation formula for the kth feature is as follows:

[0030]

[0031] In the above formula, cov(·) and dev(·) represent the covariance and standard deviation respectively, the value range of α is [0,1], and the regulatory label γ formula is as follows:

[0032]

[0033] In the above formula, the total number of columns of regulatory labels γ is r;

[0034] Perform Softmax operation on the above data sample O and supervision label γ to obtain and In obtaining and After that, the loss value l is defined by the distribution divergence metric, using the Jensen-Shannon divergence, and the formula is as follows:

[0035]

[0036] In the above formula, the overall loss function formula for scale-aligned distribution learning is as follows:

[0037]

[0038] In the above formula, O x and γ x represents the supervisory signal created from the original data instance d.

[0039] Preferably, in step S4, the test set is input into the trained anomaly detection network, and the training of the anomaly detection network and the degree of anomaly of the input data are measured by the loss function, and the process is as follows:

[0040] For a test data instance d, the scale-aligned distribution learning framework creates a transformed data sample O and a corresponding supervision label γ. The formula for the anomaly score is as follows:

[0041]

[0042] You can set the corresponding anomaly score threshold according to your needs to achieve unsupervised anomaly detection.

[0043] Preferably, in step S5, the process of deploying the anomaly detection network in a real working environment is as follows:

[0044] The model is fine-tuned using images generated from actual working conditions. The model formula is: The formula for the fine-tuning process is as follows:

[0045]

[0046] In the above formula, θ * is the fine-tuned model parameter, L t is the transfer learning training process, y t Label the target domain samples;

[0047] In engineering practice, the diagnostic recognition effect and robustness of the trained model under different health conditions of photovoltaic images are tested, and the test set samples are selected. The model was tested and the classification diagnosis accuracy index was used to evaluate the diagnostic recognition accuracy and robustness of the model in the photovoltaic panel EL image dataset. The formula of the diagnosis accuracy index is as follows:

[0048]

[0049] In the above formula, is the model prediction result, y i is the true health status label of the sample.

[0050] Therefore, the present invention adopts the above-mentioned unsupervised learning method for intelligent health monitoring of photovoltaic systems, which has the following advantages:

[0051] (1) In the present invention, by establishing an anomaly monitoring network training platform, the corresponding anomaly detection network is trained to form a unified operation, and the anomaly detection network can be quickly trained according to different detection objects.

[0052] (2) In the present invention, feature modeling of electroluminescent (EL) images of photovoltaic panels is achieved through a feature extraction framework based on multi-head linear attention. The linear attention mechanism is used in the feature extraction process, which greatly reduces the computational requirements of this stage.

[0053] (3) In the present invention, an unsupervised anomaly detection framework based on scale learning is designed. The core of the framework includes feature random sampling, feature dimension transformation, generation of effective supervision signals, and scale distribution alignment learning.

[0054] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The overall structure diagram of a feature extraction network based on multi-head linear attention in an unsupervised learning method for intelligent health monitoring of photovoltaic systems of the present invention;

[0056] Figure 2 It is a structural diagram of the EfficientViT module in a feature extraction network based on multi-head linear attention in an unsupervised learning method for intelligent health monitoring of photovoltaic systems in the present invention;

[0057] Figure 3 It is a flow chart of the EfficientViT module in a feature extraction network based on multi-head linear attention in an unsupervised learning method for intelligent health monitoring of photovoltaic systems in the present invention;

[0058] Figure 4 This is a structural diagram of an unsupervised anomaly detection framework in an unsupervised learning method for intelligent health monitoring of photovoltaic systems in the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. The specific model specifications need to be selected and determined according to the actual specifications of the device, and the specific selection calculation method adopts the existing technology in the field, so it will not be described in detail.

[0060] Example

[0061] The present invention provides an unsupervised learning method for intelligent health monitoring of photovoltaic systems, comprising the following steps:

[0062] S1: Collect image data, establish a training platform for anomaly detection network, collect EL images of single crystal and polycrystalline photovoltaic panels under constant lighting conditions and constant load, perform preprocessing operations on the collected EL images, and use data enhancement technology to construct training and test sets for the image data set, which are divided into training set and test set. and test set x i represents an EL image sample, y i is a label, which indicates a faulty sample or a healthy sample. In an unsupervised scenario, the training set consists of healthy samples, and the test set includes both faulty samples and healthy samples.

[0063] The training platform of the anomaly detection network includes a feature extraction network based on multi-head linear attention and an unsupervised anomaly detection framework based on scale learning. The training and learning of the anomaly detection network are carried out through the training platform of the anomaly detection network.

[0064] S2: Figure 1 , the training set is input into the feature extraction network based on multi-head linear attention. The feature extraction network based on multi-head linear attention includes five stages, namely the input stage and the subsequent four feature extraction stages. The input stage is a convolution layer module and a depth-separable convolution module. In the first and second feature extraction stages, an inverted linear bottleneck layer module is set. The step size of the inverted linear bottleneck layer module in each stage is set to 2. In the third and fourth feature extraction stages, a depth-separable convolution module and an EfficientViT module are set.

[0065] like Figure 2-Figure 3In the EfficientViT module, there are multi-scale linear attention modules for global information extraction, as well as feedforward layers and deep convolutional layers for local information extraction. The input is linearly projected to obtain the Q / K / V matrix, and the Q / K / V matrix is ​​aggregated with lightweight small-kernel convolution to obtain a multi-scale matrix. The multi-scale matrix is ​​linearly focused by ReLU, and the output is connected to the final linear projection layer for feature fusion. The linear attention calculation formula is as follows:

[0066]

[0067] In the above formula, when calculating linear attention, only and The results can be reused, so only O(N) computational cost and O(N) memory usage are required, thereby reducing the computational complexity and memory usage from quadratic to linear.

[0068] The deep features of photovoltaic EL images are modeled and image features are extracted. The results of the first, third and fourth feature extraction stages are upsampled, the number of channels and spatial dimensions are adjusted to remain consistent, and feature fusion is performed. The feature fusion results are output through an inverted linear bottleneck layer, and the output result is the image feature.

[0069] S3: The image features extracted in step S2 are converted into one-dimensional vectors and then input into the unsupervised anomaly detection framework based on scale learning, such as Figure 4 , is the flow chart of the anomaly detection framework.,Through random sampling, feature conversion, and scale calculation methods,,effective supervision signals are generated for distribution alignment learning.,The process of distribution alignment learning is as follows;

[0070] Create a scale-based supervision signal, perform random sampling on the input one-dimensional vector several times, and obtain the randomly sampled features for a specific feature subspace S i , the features of which Input them into the feature conversion function T and the scale label calculation function G respectively, and obtain the data sample O output by the feature conversion function T and the supervision label γ output by the scale label calculation function G;

[0071] The feature subspace S i Features in The input is converted into n-dimensional data by the feature conversion function T. The feature conversion function T is defined as a simple feedforward layer with random initialization. Each feature subspace corresponds to a transformation layer, and the formula is as follows:

[0072]

[0073] In the above formula, represents the v-dimensional subvector, X v ∈R n×v Represents the weight matrix, b∈R n represents the bias term;

[0074] For m randomly sampled subvectors of a data instance d, their transformation is represented by the matrix W∈R m×n Indicates that W is regarded as a single data sample for scale learning, and the formula for data sample O is as follows:

[0075]

[0076] In the above formula, r is the total number of columns of data sample O;

[0077] The characteristics Input into the scale label calculation function G, the formula is as follows:

[0078]

[0079] In the above formula, n is the representation dimension of the sub-vector, α k is the weight of the kth feature, γ is the amplification factor, α k The formula is as follows:

[0080]

[0081] In the above formula, the weight calculation formula for the kth feature is as follows:

[0082]

[0083] In the above formula, cov(·) and dev(·) represent the covariance and standard deviation respectively, the value range of α is [0,1], and the regulatory label γ formula is as follows:

[0084]

[0085] In the above formula, the total number of columns of regulatory labels γ is r;

[0086] Perform Softmax operation on the above data sample O and supervision label γ to obtain and In obtaining and After that, the loss value l is defined by the distribution divergence metric, using the Jensen-Shannon divergence, and the formula is as follows:

[0087]

[0088] In the above formula, the overall loss function formula for scale-aligned distribution learning is as follows:

[0089]

[0090] In the above formula, O x and γ x represents the supervisory signal created from the original data instance d.

[0091] Through the supervision signal, the trained anomaly detection network is obtained;

[0092] S4. Input the test set into the trained anomaly detection network, complete the training of the anomaly detection network and measure the degree of anomaly of the input data through the loss function. The process is as follows:

[0093] For a test data instance d, the scale-aligned distribution learning framework creates a transformed data sample O and a corresponding supervision label γ. The formula for the anomaly score is as follows:

[0094]

[0095] You can set the corresponding anomaly score threshold according to your needs, and determine whether the rent picture is an abnormal picture based on the threshold to achieve unsupervised anomaly detection.

[0096] S5. Deploy the trained anomaly detection network in a real working environment and use the image data generated by the actual working conditions for transfer learning fine-tuning. The model formula is: The formula for the fine-tuning process is as follows:

[0097]

[0098] In the above formula, θ * is the fine-tuned model parameter, L t is the transfer learning training process, y t Label the target domain samples;

[0099] In engineering practice, the diagnostic recognition effect and robustness of the trained model under different health conditions of photovoltaic images are tested, and the test set samples are selected. The model was tested and the classification diagnosis accuracy index was used to evaluate the diagnostic recognition accuracy and robustness of the model in the photovoltaic panel EL image dataset. The formula of the diagnosis accuracy index is as follows:

[0100]

[0101] In the above formula is the model prediction result, y i The real health status labels of the samples are used to verify the fault recognition effect and robustness of the trained model in photovoltaic images under different states in engineering practice.

[0102] Therefore, the present invention adopts an unsupervised learning method for intelligent health monitoring of photovoltaic systems, proposes an anomaly detection network MLA-SDAL, extracts image features through a feature extraction network based on multi-head linear attention, realizes feature modeling of electroluminescent images of photovoltaic panels, and then sets up an unsupervised anomaly detection framework based on scale learning, through feature random sampling, feature dimension transformation, generation of effective supervision signals and scale distribution alignment learning. Robust anomaly detection is achieved by evaluating the consistency between input data distribution and model output distribution.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. An unsupervised learning method for intelligent health monitoring of photovoltaic systems, characterized by: The following steps are involved: S1: Collect image data, establish a training platform for anomaly detection network, collect EL images of monocrystalline and polycrystalline photovoltaic panels under constant lighting conditions and constant load, pre-process the collected EL images, and divide the image data set into training set and test set x i represents an EL image sample, y i is a label, which indicates a faulty sample or a healthy sample. In an unsupervised scenario, the training set consists of healthy samples, and the test set includes both faulty samples and healthy samples. The training platform of the anomaly detection network includes a multi-head linear attention-based feature extraction network and a scale-learning-based unsupervised anomaly detection framework; S2: Input the training set into the feature extraction network based on multi-head linear attention to model the deep features of photovoltaic EL images and extract image features; S3: The image features extracted in step S2 are converted into one-dimensional vectors and then input into the unsupervised anomaly detection framework based on scale learning. Through random sampling, feature conversion, and scale calculation methods, effective supervision signals are generated for distribution alignment learning to obtain a trained anomaly detection network. The process of performing distribution alignment learning is as follows; Create a scale-based supervision signal, perform random sampling on the input one-dimensional vector several times, and obtain the randomly sampled features for a specific feature subspace S i , the features of which Input them into the feature conversion function T and the scale label calculation function G respectively, and obtain the data sample O output by the feature conversion function T and the supervision label γ output by the scale label calculation function G; The feature subspace S i Features in The input is converted into n-dimensional data by the feature conversion function T. The feature conversion function T is defined as a simple feedforward layer with random initialization. Each feature subspace corresponds to a transformation layer, and the formula is as follows: In the above formula, represents the v-dimensional subvector, X v ∈R n×v Represents the weight matrix, b∈R n represents the bias term; For m randomly sampled subvectors of a data instance d, their transformation is represented by the matrix W∈R m×n Indicates that W is regarded as a single data sample for scale learning, and the formula for data sample O is as follows: In the above formula, r is the total number of columns of data sample O; The characteristics Input into the scale label calculation function G, the formula is as follows: In the above formula, n is the representation dimension of the sub-vector, α k is the weight of the kth feature, γ is the amplification factor, α k The formula is as follows: In the above formula, the weight calculation formula for the kth feature is as follows: In the above formula, cov(·) and dev(·) represent the covariance and standard deviation respectively, the value range of α is [0,1], and the regulatory label γ formula is as follows: In the above formula, the total number of columns of regulatory labels γ is r; Perform Softmax operation on the above data sample O and supervision label γ to obtain and In obtaining and After that, the loss value Defined by the distribution divergence measure, using the Jensen-Shannon divergence, the formula is as follows: In the above formula, the overall loss function formula for scale-aligned distribution learning is as follows: In the above formula, O x and γ x represents the supervisory signal created by the original data instance d; S4, input the test set into the trained anomaly detection network, calculate the anomaly score, and perform unsupervised anomaly detection on the test set images by setting the anomaly judgment threshold; S5. Deploy the trained anomaly detection network in a real working environment, and use the image data generated by actual working conditions for transfer learning fine-tuning. In engineering practice, verify the fault recognition effect and robustness of the trained model for photovoltaic images under different states.

2. The unsupervised learning method for intelligent health monitoring of photovoltaic systems according to claim 1, characterized in that: In step S1, data enhancement technology is used to construct a training set and a test set.

3. The unsupervised learning method for photovoltaic system intelligent health monitoring according to claim 1, characterized in that: In the step S2, the feature extraction network based on multi-head linear attention includes five stages, namely an input stage and four subsequent feature extraction stages, the input stage is a convolution layer module and a depthwise separable convolution module, in the first and second feature extraction stages, an inverted linear bottleneck layer module is provided, the step size of the inverted linear bottleneck layer module in each stage is set to 2, and in the third and fourth feature extraction stages, a depthwise separable convolution module and an EfficientViT module are provided; In the EfficientViT module, there is a multi-scale linear attention module for global information extraction, as well as a feedforward layer and a deep convolution layer for local information extraction. The input is linearly projected to obtain the Q / K / V matrix, and the Q / K / V matrix is ​​aggregated with a lightweight small-kernel convolution to obtain a multi-scale matrix. The multi-scale matrix is ​​linearly focused on by ReLU, and the output is connected to the final linear projection layer for feature fusion. The linear attention calculation formula is as follows: In the result of the above formula, when calculating linear attention, the calculation and Upsample the results of the first, third, and fourth feature extraction stages, adjust the number of channels and spatial dimensions to keep them consistent, and perform feature fusion. The result is output through the inverted linear bottleneck layer, and the output result is the image feature.

4. The unsupervised learning method for photovoltaic system intelligent health monitoring according to claim 3 is characterized in that: In step S4, the test set is input into the trained anomaly detection network, and the training of the anomaly detection network and the degree of anomaly of the input data are measured through the loss function. The process is as follows: For a test data instance d, the scale-aligned distribution learning framework creates a transformed data sample O and a corresponding supervision label γ. The formula for the anomaly score is as follows: You can set the corresponding anomaly score threshold according to your needs to achieve unsupervised anomaly detection.

5. The unsupervised learning method for intelligent health monitoring of photovoltaic systems according to claim 4, characterized in that: In step S5, the process of deploying the anomaly detection network in a real working environment is as follows: The model is fine-tuned using images generated from actual working conditions. The model formula is: The formula for the fine-tuning process is as follows: In the above formula, θ * is the fine-tuned model parameter, L t is the transfer learning training process, y t Label the target domain samples; In engineering practice, the diagnostic recognition effect and robustness of the trained model in photovoltaic images under different health conditions are tested, and the test set samples are selected. The model was tested and the classification diagnosis accuracy index was used to evaluate the diagnostic recognition accuracy and robustness of the model in the photovoltaic panel EL image dataset. The formula of the diagnosis accuracy index is as follows: In the above formula, is the model prediction result, y i is the true health status label of the sample.

Images